Source-linked AI summary
Alaya-EVOKE: From Linear-Scaling Supervision to Endless World
Yuanyang Yin, Gongxuan Wang, Yifan Zhan, Chuanhao Li, Kaipeng Zhang, Feng Zhao
TL;DR
Interactive world models struggle to combine persistent memory, responsive interaction, and long-horizon generation without unbounded context or sacrificing few-step capability. Evoke externalizes geometric state and trains a long-horizon teacher to supervise a three-step student, achieving state-of-the-art WBench performance, competitive VBench results, and stable hour-scale generation with bounded computational cost.
Problem
Interactive world models still lack a way to combine persistent world state, responsive control, and long-horizon generation without ever-growing denoiser context or limited few-step capability.
Method
Evoke uses a camera-indexed external world state bank and a chunk-wise sparse-attention teacher with long-horizon distribution matching to train a three-step, CFG-free student.
Results
2.11 s per 1.5 s chunk at 384 × 640 on one H200; Evoke achieves state-of-the-art WBench performance while remaining competitive on VBench-2.0 and VBench-Long.
Takeaways & Limitations
Evoke supports continuously evolving interactive generation with bounded per-step state and context as sessions grow.
Takeaways & Limitations
Persistent recall is limited to retained geometric coverage, and comparisons use unmatched sampling steps because Evoke uses three steps while peers use many-step samplers.
Abstract
from arXiv · showhide
Interactive world models must support persistent memory, responsive interaction, and long-horizon generation, yet these requirements place conflicting demands on the model. Maintaining history in the denoiser context or key-value cache incurs growing cost, forcing a trade-off between session length and retained memory, while low-latency interaction relies on few-step generation whose capabilities are bounded by its teacher. Evoke addresses both limitations by externalizing persistent world state and redesigning the teacher for long-horizon interactive generation. Scene geometry is maintained in an external, camera-indexed world state bank, from which only view-relevant information is retrieved, keeping the denoiser context bounded as the session grows. Rather than treating the teacher as a fixed generator, we design it for long-horizon supervision: its sparse attention combines chunk-wise grouping, retrieval of selected distant frames, and a linear-attention global state, yielding linear growth in memory and compute while enabling supervision over long horizons. Such supervision exposes content drift that stays locally plausible within short windows, while per-chunk conditioning enables prompt changes and event control throughout the sequence. A 30-second distribution-matching objective, applied under self-forced rollouts, transfers both capabilities to a three-step student that uses no classifier-free guidance, improving resistance to long-term drift while preserving responsive conditioning. With bounded context and recurrent external memory, Evoke supports open-ended, continuously evolving generation; on a single H200 at $384\times 640$, each $1.5\,\mathrm{s}$ chunk is generated in $2.11\,\mathrm{s}$. As a three-step world model, Evoke achieves state-of-the-art performance on WBench while remaining competitive on VBench-Long and VBench-2.0.
1 Introduction
Evoke addresses the tension between persistent memory, responsive interaction, and long-horizon generation by externalizing world state and redesigning teacher supervision. Its bounded recurrent process and three-step student enable efficient, coherent, open-ended generation with competitive benchmark performance.
- Motivation: Interactive world models must preserve persistent world state, respond promptly to camera and text controls, and generate sessions lasting seconds to hours.Retaining history in denoiser context or accumulated key-value caches makes denoising cost grow with session length, while bounded alternatives discard information or face finite context budgets.
- External World State: Evoke maintains scene geometry in an external, bounded world state bank indexed by camera pose, keeping persistent state outside the denoiser.Previously observed surfaces re-entering the camera frustum are rendered into the current view as pixel-space conditioning.
- Long-Horizon Supervision: Evoke redesigns the teacher with chunk-wise sparse attention that combines bounded local context, selected distant frames, and a linear-attention global state.This reduces attention growth from quadratic to linear with sequence length and enables independent text conditioning for each chunk.
- Open-Ended Generation: Evoke reuses the same local positional layout for every generated chunk, so session duration does not expand the positional range encountered during training.Persistent world state is accessed through camera pose rather than temporal position, supporting a recurrent process independent of elapsed time.
- Results: 2.11 s generates a 1.5 s chunk on a single H200 at 384 × 640 with the world state bank enabled, while Evoke achieves state-of-the-art performance on WBench.The three-step student also demonstrates coherent hour-scale generation and remains competitive with many-step systems on VBench-2.0 and VBench-Long.
2 Related work
Interactive world models differ in how they represent and retain session information, including denoiser-side history, geometry-based conditioning, and teacher–student inference. Evoke combines these directions while relocating persistent state to an explicit external world state bank with bounded access, keeping denoiser context independent of session duration.
- Interactive world models: Interactive video world models differ fundamentally in how they represent and retain information across a running session.Related approaches include denoiser-internal history through additional context frames or an expanding key-value cache.
- Bounded retention: With bounded retention, per-step cost does not grow with session length.The camera pose reads the world state bank into the target view, and the emitted chunk updates the store and history.
- Teacher–student inference: Recent work examines the mismatch between bidirectional teacher supervision and causal student inference.
- Evoke’s design focus: Evoke builds on geometry-based conditioning while shifting persistent state from individual mechanisms to an explicit external world state bank.Previously observed scene content is rendered into the current view, while the bank has bounded read, write, and eviction operations.
3 Evoke
Evoke enables long-horizon interactive generation through a bounded recurrent student with camera-indexed external world state and a teacher designed for long-horizon supervision. Fixed-budget recurrence keeps per-step computation independent of session duration while preserving persistent scene information and supporting changing instructions.
- Bounded recurrent generation: Evoke’s three-step student generates each video chunk from bounded local history, retrieved view-aligned geometry, and a per-chunk text condition.The recurrent formulation avoids increasing context length or positional span as the session grows.
- Externalized world state: At each step, Evoke reads geometry relevant to the camera trajectory, generates the next chunk, and writes the observation back into the external world state bank.The interface is defined by rk = Read(Mk, Pk), xk ∼pθ(· | rk, hk, ck), and Mk+1 = Write(Mk, xk, Pk).
- Bounded recurrent generation: Fixed budgets for local history and world state mean that extending a session increases only recurrent calls, not an individual call’s context length, positional span, or computational footprint.Hour-scale inference retains 2160 pixel frames, corresponding to 90 s of geometry, with at most 720 active source frames.
- Long-horizon supervision: Evoke treats the teacher as a model explicitly designed to supervise long-horizon interactive generation, because the student’s acquired behavior is limited by the supervision it receives.The teacher must expose long windows and time-varying conditioning encountered during deployment.
- Long-horizon supervision: Chunk-wise sparse attention makes supervision over long, time-varying clips affordable by giving each query chunk fixed key sources, linear growth with sequence length, and access to its covering text segment.The sampled timestep selects the high- or low-noise expert, and long-rollout distribution matching follows camera-control training and few-step distillation.
4 Experiments
Experiments show that Evoke combines strong interactive and general video quality with stable, bounded-cost long-session generation. Ablations further link long-horizon teacher supervision to reduced drift and demonstrate complementary control through persistent geometric memory and timed text conditioning.
- Interactive world-model performance: Evoke leads few-step systems on WBench’s Video Quality, Setting, and Physical group averages while matching the strongest Consistency result.The evaluation uses the 158-case navigation split.
- General video quality: 66.77 on VBench-2.0 and 85.11 on VBench-Long show competitive general video quality despite three-step, classifier-free-guidance-free evaluation.Peers use many-step default samplers, so comparisons are not step-matched.
- Long-session stability and efficiency: Eight 65.5-minute sessions show little post-transient photometric drift, while fixed retention keeps recurrent-step state and computation stable as sessions lengthen.The evaluated rollouts contain 2,619 chunks each, with 90 s of retained observations.
- Long-session stability and efficiency: 2.11 s per 1.5 s chunk on one H200 demonstrates responsive generation under the reported 384 × 640 runtime protocol.Measurements use the full VAE decoder without key-value caching, compilation, quantization, or a distilled decoder.
- Teacher supervision: Long-horizon teacher supervision transfers substantially stronger photometric stability to the few-step student than short-horizon supervision.Scoring sweeps find no consistent sharp detectability threshold beyond a short horizon, so the benefit is not explained by window length alone.
- Memory and timed control: 67% of unanchored timed text clauses are realized versus 4% for clauses requiring anchored-geometry replacement, while pose-addressed recall improves when retention covers the revisit.Text controls content free to evolve; persistent geometry resists overwriting anchored observations, and geometric-conditioning cost depends on retrieved visible coverage rather than session duration.
5 Conclusion
Evoke is a three-step video world model that combines external geometric memory with long-horizon teacher supervision for interactive generation. Its conclusion highlights bounded recurrent context and state, while identifying fine-grained consistency and dynamic world-state modeling as open challenges.
- Contributions: Evoke is a three-step video world model for long-horizon interactive generation.Its design decouples persistent world state from ever-growing model context through external geometric memory.
- Contributions: External geometric memory keeps the state and context of each recurrent generation step bounded as sessions grow.The model decouples persistent world state from the model’s ever-growing context while enabling continuous camera and text control.
- Contributions: A teacher redesigned for long-horizon generation provides extended supervision for Evoke’s few-step student.The teacher supports the three-step world model formulation described in the conclusion.
- Limitations and future work: Fine-grained consistency of object identity, appearance, and local details remains limited because the geometric world state primarily preserves coarse scene structure.The conclusion suggests richer object-level or semantic world representations as a potential way to improve long-term identity consistency.
- Limitations and future work: Future work should model dynamic world state, including object motion, state transitions, and their long-term evolution.The conclusion distinguishes dynamic state from static geometry as an open direction.
Appendix · A Training curves
The released Evoke student is produced through two stable distillation stages: long-distill against the Evoke Teacher followed by a short post-distill continuation. Training remains stable across bounded DMD, critic convergence, plateauing normalized gradients, and bounded gradient norms.
- A Training curves: Two distillation stages produce the released Evoke student: long-distill against the Evoke Teacher, followed by a short post-distill continuation.The long-distill stage uses 6 × 8 GPUs.
- A Training curves: 1981 steps are used for long-distill with the Evoke Teacher on 6 × 8 GPUs.The figure distinguishes per-step traces from their 120-step exponential moving average.
- A Training curves: DMD stays bounded throughout both distillation stages.The optimization record shows stable DMD behavior in long-distill and post-distill continuation.
- A Training curves: The critic converges during the two-stage distillation process.Critic convergence is reported as one of the stable training indicators.
- A Training curves: The normalized gradient entering the loss plateaus, and release occurs at its onset.The plateau marks the release point for the student.
- A Training curves: Gradient norms stay bounded across 48 GPUs and eight scheduler restarts.This behavior is reported for the full optimization record.
B Cost of the geometric path
The geometric path is a substantial recurrent-step cost, adding 1.84 s per chunk while scaling with warp coverage. Its denoiser impact is confined to coarsest-stage token admission; later stages remain unchanged.
- Geometric-path cost: 38% of the denoiser and 93% of a recurrent step are consumed by the geometric path when it is included.Measured per chunk on one H200 at 384×640 with 3 denoising steps and classifier-free guidance off.
- Geometric-path cost: 1.84 s per chunk is added by geometry, with cost scaling according to warp coverage.Unobserved space costs less, precisely where memory provides less coverage.
- Geometric-path cost: Warp rendering and coarsest-stage token admission are separable costs, while stages 1 and 2 remain unchanged throughout.The figure attributes denoiser cost to tokens rather than geometry itself.
C Long-Horizon Generation and Teacher Ablations
Long-session evaluation combines extended CFG-free rollouts with photometric and self-consistency measures, while ablations show that long-horizon teacher supervision improves photometric stability but not measured content separation, sharpness, or drift. Additional analyses find no simple scoring-window threshold and identify per-chunk conditioning as the mechanism for timed prompt changes without establishing a teacher-horizon advantage in event realization.
- Long-session evaluation: 2,619 recurrent steps and 65.5 minutes define each of eight continuous long-session rollouts, with 94,281 generated frames at 384 × 640 on one H200.Each chunk uses three CFG-free evaluations, the world state bank retains 90 s of observations, and 2.11 s denotes diffusion wall-clock latency rather than end-to-end execution.
- Long-session evaluation: Photometric statistics track low-level appearance changes, whereas the content descriptor measures self-consistency against each session’s opening segment rather than reference-based identity.Real-video controls can decorrelate substantially as the camera traverses new content, so descriptor change is not itself evidence of model degradation.
- Teacher ablations: Long-horizon teacher supervision improves photometric stability over extended rollouts, but students do not significantly separate on content descriptor, sharpness, or final-checkpoint drift advantage.The comparison uses few-step students distilled with matched recipes from the short-horizon and Evoke teachers.
- Teacher ablations: 13 controlled perturbation conditions show little change in teacher-critic detectability beyond W = 2 chunks, with no consistent sharp threshold as W increases.Similar response curves across perturbation-to-scored-chunk distances indicate that simple window coverage is insufficient to explain the observed teacher behavior.
- Teacher ablations: Per-chunk conditioning supports mid-session prompt changes, but matched short- and long-horizon teachers do not significantly differ in event-realization rate at the current sample size.Four four-minute teacher rollouts, each driven by twelve consecutive 20 s instructions, show visible effects within every segment without resetting.
D Qualitative rollouts
The qualitative rollouts illustrate Evoke across diverse interactive settings and long, changing instruction schedules. Frames are sampled throughout each clip or instruction segment, with joystick overlays showing camera control, but the figures provide no quantitative measurements.
- Figure 12: Figure 12 covers nine sessions across four settings: egocentric action and manipulation, cinematic scenes and styles, embodied and industrial scenes, and timed events.The settings span interactive actions, visual styles, embodied or industrial environments, and time-triggered events.
- Figure 12: Each Figure 12 row samples five frames evenly across its clip, depicting rollout behavior over the full sequence rather than selected moments.Timed-event rows are longer because their instructions fire during the last third of the clip.
- Figure 11: Figure 11 shows four-minute Evoke Teacher rollouts driven by twelve instructions held for 20 s apiece, with frames selected within each instruction segment.The figure depicts four sessions and distinguishes illustrated schedule segments from those not shown.
- Figure 12: The lower-corner rings show the interactive demo’s joystick overlay and the camera command driving motion.This overlay appears on every tile in the Figure 12 sheet.
- Qualitative assessment: Both figures are qualitative and carry no measurement.The figures emphasize visualizing rollout behavior rather than reporting quantitative evaluation.
E Public leaderboards in full
Evoke is evaluated against the full public top-ten leaderboards for VBench-2.0 and VBench-Long, with per-dimension breakdowns and explicitly declared protocol deviations. It uses three-step, CFG-free sampling while peers use their default many-step samplers, and leaderboard scores are reported on comparable scales.
- Public leaderboard coverage: Tables 3–6 expand the public top-ten VBench-2.0 and VBench-Long leaderboards with Evoke and provide per-dimension scores.VBench-2.0 aggregates raw dimension scores, whereas VBench-Long min–max-normalizes before weighting.
- VBench-2.0 protocol: 4 declared deviations: VBench-2.0 uses one sample per prompt, 5.875 s clips, 640 × 384 resolution, and prompt augmentation.Diversity alone uses official raw prompts because augmentation was measured to cost 3.38 points.
- VBench-Long protocol: 3 declared deviations: VBench-Long uses one sample per prompt and 8.875 s clips rather than 10 s, with the remaining deviation declared in the protocol passage.The Evoke row is de-normalized with official constants for comparison on the leaderboard scale.
- Sampling protocol: 3 network evaluations per chunk: Evoke uses a three-stage pyramid with one step per stage and no classifier-free guidance, unlike peers’ many-step default samplers.This sampling setup applies to the VBench-2.0 comparison.
F WBench public leaderboard
Evoke is evaluated on the WBench navigation split against public systems spanning different sampling budgets. The comparison places Evoke’s three-step, guidance-free student on the public leaderboard but is not step-matched.
- Evaluation setup: Evoke is compared with eight few-step interactive systems on WBench’s navigation split.The same run is also placed in the public WBench leaderboard for that split.
- Evaluation setup: The public leaderboard ranks 30 systems by the unweighted mean of five group scores, with navigation as the sole Interaction dimension on this split.Table 7 reports the Navi split with n = 158 cases and lists the top ten by Average.
- Comparison caveat: Evoke uses three sampling steps without classifier-free guidance, whereas peers use their default samplers, so the comparison is not step-matched.The result is an evaluation of the released student rather than a leaderboard submission.